File size: 11,125 Bytes
1ea7ba6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 | """
3-Phase Trainer for SpiceFusionNet.
Phase 1 — Backbone pre-training (CE loss, image-only)
Phase 2 — Contrastive fine-tuning (SupCon loss, backbone only)
Phase 3 — Full fusion end-to-end (CE + SupCon, all branches)
"""
import time
from pathlib import Path
import torch
import torch.nn as nn
from torch.optim import AdamW
from torch.optim.lr_scheduler import CosineAnnealingLR, LinearLR, SequentialLR
from torch.utils.data import DataLoader
import config
from src.model import SpiceFusionNet, save_checkpoint
from src.losses import SupConLoss, CombinedLoss
try:
import wandb as _wandb
_WANDB = True
except ImportError:
_WANDB = False
def _log(d: dict):
if config.USE_WANDB and _WANDB:
_wandb.log(d)
def _make_scheduler(optimizer, warmup: int, total: int, min_lr: float):
return SequentialLR(
optimizer,
schedulers=[
LinearLR(optimizer, start_factor=1e-3, end_factor=1.0, total_iters=warmup),
CosineAnnealingLR(optimizer, T_max=total - warmup, eta_min=min_lr),
],
milestones=[warmup],
)
class PhaseTrainer:
def __init__(self, model: SpiceFusionNet, device: torch.device, ckpt_dir: Path):
self.model = model
self.device = device
self.ckpt_dir = ckpt_dir
ckpt_dir.mkdir(parents=True, exist_ok=True)
# ── Phase 1 ──────────────────────────────────────────────────────
def phase1(self, train_loader: DataLoader, val_loader: DataLoader) -> dict:
print("\n" + "="*60)
print(" PHASE 1 — Backbone Pre-training")
print("="*60)
model = self.model.to(self.device)
criterion = nn.CrossEntropyLoss(label_smoothing=config.P1_LABEL_SMOOTH)
optimizer = AdamW(model.parameters(), lr=config.P1_LR, weight_decay=config.P1_WEIGHT_DECAY)
scheduler = _make_scheduler(optimizer, config.P1_WARMUP, config.P1_EPOCHS, config.P1_MIN_LR)
best_acc, patience, history = 0.0, 0, {"train_loss": [], "train_acc": [], "val_loss": [], "val_acc": [], "lr": []}
for epoch in range(1, config.P1_EPOCHS + 1):
t0 = time.time()
# Train
model.train()
tl, tc, tt = 0.0, 0, 0
for imgs, tex, col, labels in train_loader:
imgs, labels = imgs.to(self.device), labels.to(self.device)
logits = model.forward_image(imgs)
loss = criterion(logits, labels)
optimizer.zero_grad(); loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), config.GRAD_CLIP)
optimizer.step()
tl += loss.item() * imgs.size(0)
tc += (logits.argmax(1) == labels).sum().item()
tt += imgs.size(0)
# Val
val_loss, val_acc = self._eval_image(model, val_loader, criterion)
scheduler.step()
lr = optimizer.param_groups[0]["lr"]
for k, v in zip(["train_loss","train_acc","val_loss","val_acc","lr"],
[tl/tt, tc/tt, val_loss, val_acc, lr]):
history[k].append(v)
print(f"P1 Ep {epoch:03d}/{config.P1_EPOCHS} | "
f"tr_loss {tl/tt:.4f} tr_acc {tc/tt:.4f} | "
f"val_loss {val_loss:.4f} val_acc {val_acc:.4f} | "
f"lr {lr:.2e} | {time.time()-t0:.1f}s")
_log({"p1/train_loss": tl/tt, "p1/val_acc": val_acc, "p1/lr": lr})
if val_acc > best_acc:
best_acc, patience = val_acc, 0
save_checkpoint(self.ckpt_dir/"p1_best.pth", model, optimizer, epoch, best_acc, history)
print(f" --> P1 best: {best_acc:.4f}")
else:
patience += 1
if patience >= config.PATIENCE:
print(f" Early stop at epoch {epoch}")
break
save_checkpoint(self.ckpt_dir/"p1_last.pth", model, optimizer, epoch, best_acc, history)
print(f"\nPhase 1 complete. Best val acc: {best_acc:.4f}")
return history
# ── Phase 2 ──────────────────────────────────────────────────────
def phase2(self, train_loader: DataLoader) -> None:
print("\n" + "="*60)
print(" PHASE 2 — Contrastive Fine-tuning (SupCon)")
print("="*60)
# Load best P1 weights
p1_ckpt = self.ckpt_dir / "p1_best.pth"
if p1_ckpt.exists():
ckpt = torch.load(p1_ckpt, map_location=self.device)
self.model.load_state_dict(ckpt["model_state"])
print(f" Loaded P1 best checkpoint (val acc: {ckpt.get('best_val_acc',0):.4f})")
model = self.model.to(self.device)
supcon = SupConLoss()
# Only train backbone + proj_head; freeze branches and heads
for name, p in model.named_parameters():
p.requires_grad = any(k in name for k in ("backbone", "proj_head"))
trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(f" Trainable params: {trainable:,}")
optimizer = AdamW(
filter(lambda p: p.requires_grad, model.parameters()),
lr=config.P2_LR,
)
for epoch in range(1, config.P2_EPOCHS + 1):
t0 = time.time()
model.train()
total_loss, n = 0.0, 0
for imgs, tex, col, labels in train_loader:
imgs, labels = imgs.to(self.device), labels.to(self.device)
proj = model.forward_contrastive(imgs)
loss = supcon(proj, labels)
optimizer.zero_grad(); loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), config.GRAD_CLIP)
optimizer.step()
total_loss += loss.item() * imgs.size(0)
n += imgs.size(0)
avg = total_loss / n
print(f"P2 Ep {epoch:02d}/{config.P2_EPOCHS} | SupCon loss {avg:.4f} | {time.time()-t0:.1f}s")
_log({"p2/supcon_loss": avg})
# Re-enable all params for Phase 3
for p in model.parameters():
p.requires_grad = True
save_checkpoint(self.ckpt_dir/"p2_last.pth", model, optimizer, epoch, 0.0, {})
print("Phase 2 complete.")
# ── Phase 3 ──────────────────────────────────────────────────────
def phase3(self, train_loader: DataLoader, val_loader: DataLoader) -> dict:
print("\n" + "="*60)
print(" PHASE 3 — Full Fusion End-to-end Training")
print("="*60)
p2_ckpt = self.ckpt_dir / "p2_last.pth"
if p2_ckpt.exists():
ckpt = torch.load(p2_ckpt, map_location=self.device)
self.model.load_state_dict(ckpt["model_state"])
print(" Loaded P2 checkpoint.")
model = self.model.to(self.device)
loss_fn = CombinedLoss()
optimizer = AdamW(model.parameters(), lr=config.P3_LR, weight_decay=config.P3_WEIGHT_DECAY)
scheduler = CosineAnnealingLR(optimizer, T_max=config.P3_EPOCHS, eta_min=1e-7)
best_acc, patience, history = 0.0, 0, {"train_loss": [], "train_acc": [], "val_loss": [], "val_acc": [], "lr": []}
for epoch in range(1, config.P3_EPOCHS + 1):
t0 = time.time()
model.train()
tl, tc, tt = 0.0, 0, 0
for imgs, tex, col, labels in train_loader:
imgs, tex, col, labels = (
imgs.to(self.device), tex.to(self.device),
col.to(self.device), labels.to(self.device),
)
logits, proj = model.forward_fusion(imgs, tex, col)
loss = loss_fn(logits, proj, labels)
optimizer.zero_grad(); loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), config.GRAD_CLIP)
optimizer.step()
tl += loss.item() * imgs.size(0)
tc += (logits.argmax(1) == labels).sum().item()
tt += imgs.size(0)
val_loss, val_acc = self._eval_fusion(model, val_loader)
scheduler.step()
lr = optimizer.param_groups[0]["lr"]
for k, v in zip(["train_loss","train_acc","val_loss","val_acc","lr"],
[tl/tt, tc/tt, val_loss, val_acc, lr]):
history[k].append(v)
print(f"P3 Ep {epoch:02d}/{config.P3_EPOCHS} | "
f"tr_loss {tl/tt:.4f} tr_acc {tc/tt:.4f} | "
f"val_loss {val_loss:.4f} val_acc {val_acc:.4f} | "
f"lr {lr:.2e} | {time.time()-t0:.1f}s")
_log({"p3/val_acc": val_acc})
if val_acc > best_acc:
best_acc, patience = val_acc, 0
save_checkpoint(self.ckpt_dir/"best.pth", model, optimizer, epoch, best_acc, history)
print(f" --> Best: {best_acc:.4f}")
else:
patience += 1
if patience >= config.PATIENCE:
print(f" Early stop at epoch {epoch}")
break
save_checkpoint(self.ckpt_dir/"last.pth", model, optimizer, epoch, best_acc, history)
print(f"\nPhase 3 complete. Best val acc: {best_acc:.4f}")
return history
# ── Helpers ───────────────────────────────────────────────────────
@torch.no_grad()
def _eval_image(self, model, loader, criterion):
model.eval()
tl, tc, tt = 0.0, 0, 0
for imgs, tex, col, labels in loader:
imgs, labels = imgs.to(self.device), labels.to(self.device)
logits = model.forward_image(imgs)
tl += criterion(logits, labels).item() * imgs.size(0)
tc += (logits.argmax(1) == labels).sum().item()
tt += imgs.size(0)
return tl / tt, tc / tt
@torch.no_grad()
def _eval_fusion(self, model, loader):
model.eval()
criterion = nn.CrossEntropyLoss()
tl, tc, tt = 0.0, 0, 0
for imgs, tex, col, labels in loader:
imgs, tex, col, labels = (
imgs.to(self.device), tex.to(self.device),
col.to(self.device), labels.to(self.device),
)
logits, _ = model.forward_fusion(imgs, tex, col)
tl += criterion(logits, labels).item() * imgs.size(0)
tc += (logits.argmax(1) == labels).sum().item()
tt += imgs.size(0)
return tl / tt, tc / tt
|